CC Llama Vision Advanced Options

A ComfyUI node in llama.cpp with 22 inputs and 1 output.

By cicalooo·Created 2 months ago·Updated about a month ago· 1

advanced_options

◄temperature0.90►

◄top_p0.90►

◄top_k64►

◄min_p0.00►

◄repeat_penalty1.00►

◄presence_penalty0.00►

◄frequency_penalty0.00►

◄seed0►

◄disable_thinkingtrue►

◄stop_sequences►

◄n_gpu_layers99►

◄ctx_size8192►

◄threads0►

◄threads_batch0►

◄extra_server_args►

◄keep_server_alivefalse►

◄idle_timeout_s60►

◄force_restartfalse►

◄startup_timeout_s60►

◄request_timeout_s300►

◄debugtrue►

◄server_log_path/root/Documents/ComfyUI/llama_server_debug.log►

Categoryllama.cpp

Inputs (22)

Name	Type	Default	Description
temperature	FLOAT	0.900–2	Sampling temperature. Higher = more random/creative, lower = more deterministic/focused.
top_p	FLOAT	0.900–1	Nucleus sampling threshold — only tokens within this cumulative probability mass are considered.
top_k	INT	640–1000	Only the top K most likely tokens are considered at each step. 0 disables this filter.
min_p	FLOAT	0.000–1	Minimum probability (relative to the top token) a token must have to be considered.
repeat_penalty	FLOAT	1.000–2	Penalty applied to tokens that have already appeared, to discourage repetition.
presence_penalty	FLOAT	0.00-2–2	Penalty applied per unique token already present in the output, encouraging new topics.
frequency_penalty	FLOAT	0.00-2–2	Penalty scaled by how often a token has already appeared, discouraging repetition.
seed	INT	0-1–18446744073709550000	Random seed for generation. -1 for random each run; a fixed value makes output reproducible.
disable_thinking	BOOLEAN	true	If enabled, disables the model's internal 'thinking'/reasoning mode (for models that support it) so it responds directly.
stop_sequences	STRING		One stop string per line; generation halts early if any of these strings are produced.
n_gpu_layers	INT	990–200	Number of model layers to offload to GPU. Higher = faster but more VRAM; set to 0 for CPU-only.
ctx_size	INT	8192512–131072	Context window size (tokens) for the server. Larger allows longer prompts/images but uses more VRAM/RAM.
threads	INT	00–256	Number of CPU threads for generation. 0 = let llama-server auto-detect.
threads_batch	INT	00–256	Number of CPU threads for batch/prompt processing. 0 = let llama-server auto-detect.
extra_server_args	STRING		Additional raw command-line arguments passed through to llama-server (advanced use).
keep_server_alive	BOOLEAN	false	Keep llama-server running after this node finishes so future runs can reuse it instead of restarting (faster, but keeps VRAM occupied).
idle_timeout_s	INT	600–3600	If a kept-alive server has gone unused for this many seconds, it is killed to free VRAM. The check runs when a CC Llama Vision node next executes (there is no background timer). 0 disables auto-unload.
force_restart	BOOLEAN	false	Kill and restart any matching existing server before running, even if one is already alive and healthy.
startup_timeout_s	INT	605–300	How many seconds to wait for llama-server to report healthy before giving up on startup.
request_timeout_s	INT	30030–1800	How many seconds to wait for a response to the captioning request before timing out.
debug	BOOLEAN	true	Print the raw JSON response from llama-server to the console for troubleshooting.
server_log_path	STRING	/root/Documents/ComfyUI/llama_server_debug.log	File path where llama-server's stdout/stderr log will be written — check this file if startup fails.

Outputs (1)

Name	Type	Description
advanced_options	LLAMA_VISION_OPTS	—